GeoBoreNet: RGB–Geometry Fusion for Cross-Scene Borehole Detection

Borehole detection in textured quarry reconstructions is challenging because shadows, rocks, and fractured ground can resemble visible surface openings. We introduce GeoBoreNet, which augments RGB with four pixel-aligned inputs derived from reconstructed surface height: local residual relief, spatial gradient magnitude, spatial Laplacian response, and mesh support. Plane removal and resolution-adapted local-background subtraction encode surface morphology, and an expanded first convolution incorporates these channels into standard detectors. Experiments evaluate DINO-R50 and YOLO11x across 11 source scenes using scene-disjoint leave-one-scene-out testing, with five paired training seeds for DINO-R50. In pooled cross-fold evaluation, adding geometry increases AP from 0.483 to 0.548 for DINO-R50 and from 0.385 to 0.450 for YOLO11x, an absolute gain of 0.065 for each detector. Component and local-background scale studies further assess the geometry representation. DINO-R50 source-object-unique AP increases from 0.448 to 0.493 at cross-patch NMS IoU 0.4. In evaluated search regions containing zero-object tiles, AP increases from 0.380 to 0.430 for DINO-R50 and from 0.300 to 0.360 for YOLO11x. These results show that local surface morphology complements RGB for two-dimensional borehole detection in the evaluated scenes.